Add camera and actuator corroboration v0.4

.github/workflows/image-goal.yml

@@ -23,6 +23,8 @@ jobs:
           tests.test_image_goal_freeze tests.test_image_goal_scenes
           tests.test_image_goal_report tests.test_local_goal_verifier
           tests.test_local_goal_benchmark tests.test_local_goal_scenes
-          tests.test_local_goal_verifier_report -v
+          tests.test_local_goal_verifier_report tests.test_sensor_goal_observers
+          tests.test_sensor_goal_verifier tests.test_sensor_goal_benchmark
+          tests.test_sensor_goal_scenes tests.test_sensor_goal_verifier_report -v
       - name: Preserve published evidence and existing Space files
         run: python -m unittest discover -s publishing/huggingface -p 'test_publish*goal*.py' -v

README.md

@@ -20,6 +20,18 @@ This is measured-state verification, not learned vision or a robot controller.
 [Research](https://research.dvidia.org/papers/local-goal-verifier-v03/) ·
 [Run locally](docs/local-goal-verifier.md).
 
+**[Independent sensor corroboration v0.4](benchmarks/sensor-goal-verifier-v04/README.md).**
+RGB and PiPER drive checks prevent **69/112 deliberately false completion
+readings**, while **43 remain** through orientation, identity, support, sampling
+and replay blind spots. The conservative gate also abstains on **73/80 baseline
+true completions**, including all 64 injected sensor-fault controls. Processed
+path latency is **22.1 ms median / 31.5 ms p95**, excluding the 200 ms evidence
+window and rendering. This is a paired synthetic benchmark, not an independent
+physical-sensor or hardware result.
+[Inspect the cases](https://dvidia-robot-simulation-lab.static.hf.space/image-goals/v04/index.html) ·
+[Research](https://research.dvidia.org/papers/sensor-corroboration-v04/) ·
+[Protocol](docs/sensor-goal-verifier.md).
+
 **[Image-goal verification: measured Jev results](benchmarks/image-goal-20261019/README.md).**
 32 native PiPER fixtures test missing targets, wrong arrangements, unreleased
 contact and moving shoes that render identically to settled shoes. The runnable

benchmarks/sensor-goal-verifier-v04/README.md

@@ -0,0 +1,114 @@
+# Independent camera and actuator corroboration v0.4
+
+This release adds a CPU RGB observer and PiPER drive-channel checks to the v0.3
+physical-goal verifier. The checks can withhold a `complete` verdict when the
+visible object position/motion conflicts with the primary packet, or when the
+gripper drive does not corroborate release. They do not authenticate sensors,
+identify arbitrary objects, reconstruct depth or control a robot.
+
+## Fresh paired result
+
+The frozen run contains 272 cases across eight layouts: 96 native states, 80
+primary-packet forgeries, 64 auxiliary-channel faults and 32 cross-channel
+controls. Native generation completed with **0/96 invalid cases**. The v0.3
+baseline and v0.4 gate received the exact same primary packets.
+
+| Measure | v0.3 primary only | v0.4 camera + drive |
+| --- | ---: | ---: |
+| Scheduled / evaluated | 272 / 272 | 272 / 272 |
+| Complete verdicts | 192 | 50 |
+| Physically true completions recognized | 80/80 | 7/80 |
+| Physically false completions | 112/192 (58.3%) | 43/192 (22.4%) |
+| Baseline false completions prevented | — | 69/112 (61.6%) |
+| Baseline true completions lost to abstention | — | 73/80 (91.3%) |
+| Unknown verdicts | 0/272 | 142/272 (52.2%) |
+
+The true-completion denominator includes 64 repeated sensor-fault controls whose
+physical scenes are complete but whose auxiliary evidence is deliberately
+missing, stale, gapped or malformed. All 64 correctly become `unknown`. In the
+96-case native cohort, seven of eight visible settled scenes are recognized;
+all eight occluded scenes abstain, and one nominally visible target is obscured
+by the simulated arm. The gate is conservative and has low positive coverage.
+
+| Cohort | Cases | Primary complete | Fused complete | Fused unknown | Physical false completions after fusion |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| Native scenes | 96 | 16 | 7 | 9 | 0/80 negatives |
+| Primary forgeries | 80 | 80 | 29 | 51 | 29/80 |
+| Sensor faults | 64 | 64 | 0 | 64 | Not applicable: all are physical positives |
+| Cross-channel controls | 32 | 32 | 14 | 18 | 14/32 |
+
+The 43 residual physical false completions are retained by cause:
+
+| Control | Residual false completions |
+| --- | ---: |
+| Final-instant motion alias | 7/8 |
+| 180° yaw reversal | 8/8 |
+| Same-color visual lookalike | 7/8 |
+| Elevated support hidden from the top camera | 7/8 |
+| Refreshed camera replay | 7/8 |
+| Coherent replay of all channels | 7/8 |
+
+The four controls with seven residuals have one layout where the camera could
+not observe the green proxy and therefore abstained. Axial geometry cannot
+distinguish a half-turn. A synthetic green proxy is not authenticated object
+identity. An orthographic top view does not see elevation or table support.
+Sampled RGB and drive data miss motion between samples. All channels share the
+same simulator and can be replayed together.
+
+## Processing time
+
+| Stage | Median | p95 | Maximum |
+| --- | ---: | ---: | ---: |
+| Primary verifier | 0.42 ms | 0.58 ms | 1.15 ms |
+| RGB camera observation | 14.32 ms | 20.10 ms | 27.85 ms |
+| Actuator observation | 0.86 ms | 1.21 ms | 10.33 ms |
+| Fusion | 0.005 ms | 0.018 ms | 0.047 ms |
+| Entire processed window | 22.09 ms | 31.48 ms | 44.77 ms |
+
+The timer covers local JSON reads, PNG decoding, the v0.3 baseline, camera
+processing, actuator checks and fusion. It excludes the 200 ms observation
+history, native rendering/acquisition, networking and robot actuation. The
+optional OpenCV CPU kernel was installed; a NumPy-only fallback is available
+but was not used for this timing run. These are not hard-real-time or
+end-to-end robot response measurements. No GPU or hosted model call was used.
+
+## Run it again
+
+Use the source commit frozen in [`evidence/protocol.json`](evidence/protocol.json).
+The reported run used `90e84701ce7d5bc330ef41b02793e2ce037d702d`.
+
+```sh
+python -m pip install -e '.[arm,observations]'
+dvidia-sensor-goal generate \
+  --assets /absolute/path/to/robot-profile-upstream/assets \
+  --output benchmarks/sensor-goal-verifier-v04/evidence \
+  --seed 2026102007 --count 8 --repeats 3 --warmups 1
+dvidia-sensor-goal evaluate --run benchmarks/sensor-goal-verifier-v04/evidence
+python -m dvidia_training.sensor_goal_verifier_report \
+  --run benchmarks/sensor-goal-verifier-v04/evidence --output /tmp/sensor-goal-v04-report
+```
+
+Generation needs the pinned PiPER MuJoCo assets and an available graphics
+context. Evaluation is local CPU work. `integrity.json` covers every native
+input; the result stores one paired decision per case and all three timing
+repetitions. See [`protocol.json`](evidence/protocol.json),
+[`cases.json`](evidence/cases.json), [`benchmark.json`](evidence/benchmark.json), and
+the complete [`native evidence directory`](evidence/).
+
+## Scope
+
+The RGB path receives eleven 640×480 images at 20 ms intervals and estimates
+XY centroid and an undirected object axis from one green elongated proxy. It
+has no depth, segmentation IDs or target identity. The drive path checks one
+hundred 2 ms samples of six PiPER arm joints, signed finger positions and
+velocities, gripper command and simulated motor effort. It cross-checks
+servo/encoder consistency, release aperture and settling; motor effort is not
+object contact force. Position bounds, timestamp calibration IDs and drive
+calibration are synthetic declarations, not physical calibrations.
+
+The result establishes neither independent physical sensors nor hardware
+qualification, perception quality on ordinary objects, real-time controller
+latency, action learning, training-data transfer, or robot safety. Unknown
+means insufficient or contradictory corroboration, not physical failure.
+Original DVIDIA code is MIT. Pinned simulator code and robot assets retain the
+upstream licenses recorded in the repository.